Oracle Fusion Claw – Autonomy Arrives With A Receipt

The Work Claw Targets: Specialist Problems, Delegated By Policy

A runtime is only as interesting as the problems it is pointed at, so Greyhound Research reads the work first. Oracle has been specific about it. The release names “specialist-grade work including deep research, computation, simulation, modeling, and continuous re-planning,” and four applications carry it at launch.

In plain words, Ledger closes the books faster by matching and investigating accounting entries. Workforce Staffing builds rosters that respect skills, accreditations, availability and labor rules, and redoes them when conditions change. Shipping Consolidation combines shipments to cut freight cost without breaking delivery promises. Account Territory Growth Plan models different ways of dividing sales territories and then makes the change.

The gap Oracle is pointing at

Oracle has described the gap these fill as the work of operations research teams, in its own words teams “stacked with highly trained MIT graduates trying to optimize logistics, movement of inventory and delivery routes,” and Claw “is aimed at that type of work.” At the briefing it drew the same line between the agents it has shipped since 2025 and this. Enterprise workflows are, in Oracle’s words, “very deterministic,” and agents handle them well. The class of work Claw targets “involves massive data sets,” “a lot of constraints,” and “changing business signals and policies and rules,” the work a scheduler would otherwise take to a spreadsheet for several hours.

The division of labor inside the runtime follows from that. If the agentic application is the recipe, the runtime is the kitchen, and in this kitchen the frontier model holds the planning seat and deterministic computation holds the posting seat. The model reasons about the constraints. The system of record carries out the plan.

How far the customer lets it go

The customer chooses the level. Oracle’s deck of September 25 names three, each with a role for the people around it, and Exhibit 1 sets them out.

Oracle has described the middle level in interviews as a research mode, “where it performs the same workload and returns with the actions it proposes to take, staged for human review,” and this report uses that phrase for it. Oracle’s documentation does not yet name the levels, and Oracle has said that it expects organizations to begin with more human review.

Oracle’s documentation of September 28, 2026 states the limit on the machine’s side in one line: “Although AI can plan and recommend actions, it can’t grant itself authority to act.” Everything above that line, which outcomes qualify for full auto, at what thresholds, under whose signature, is the customer’s to decide, and the product does not decide it for them. The use case is real and well chosen. The authority to pursue it is the buyer’s to set, in the roles, approvals and policies Fusion already holds.

↑ Back to the section guide

What Is Actually Documented: A Switch, A Meter, And Fifteen Names

An evaluation starts at the distance between what a buyer can verify today and what has been announced, and Greyhound Research measures that distance here. Oracle’s release says there are “25 new Claw-powered, full auto-capable Agentic Applications available today.” Oracle’s documentation, read on October 5 and again on October 6, 2026, supports less than that sentence and more than nothing.

Three documents, one switch

Three Oracle documentation and contract documents name Fusion Claw. The Release 26D What’s New carries a feature titled “Use Oracle Fusion Claw to create autonomous outcomes with agentic applications,” delivered on September 28, 2026, one day before the release. It calls Fusion Claw “a governed and agentic runtime capability for enterprise execution” and describes setup as opening the Applications tab in AI Agent Studio, editing an application, and enabling autonomous outcomes.

The Oracle Help Center page “Fusion Claw Capabilities,” created on October 2, 2026, lists fifteen Claw-powered capabilities by name across finance, supply chain, HR and customer experience, as counted on the live page on October 6, 2026. Oracle’s Fusion Cloud Service Descriptions, revised on September 29, 2026, carry a “Fusion Claw” row in the table that prices AI consumption, with the typical uses “Deep research, Autonomous execution, Planning.” Beyond the release itself, a customer experience blog post restates it and the readiness pages point to the Help Center page. Those three documents are what Oracle had published under the name as of October 6, 2026.

Availability by release, not by announcement

Oracle’s statements on timing range wider than the release. The release says “available today.” Oracle has also said in interviews that the applications will be available in the fourth quarter, that broader availability begins in the first week of October, and that the remaining applications arrive in October. The document that settles it is Oracle’s own. Its Release Readiness pages for ERP, HCM and SCM (enterprise resource planning, human capital management and supply chain management) say “Explore the new Fusion Claw capabilities available in update 26D!”

That ties the capabilities to a quarterly update with a calendar. Oracle published its cohort rule for quarterly updates in December 2025 and has not restated it for 26D. By that rule, Release 26D reaches production environments on November 20 or December 18, 2026, or January 15, 2027, depending on a customer’s cohort. A customer running Fusion Claw in production before then is in an early-access arrangement, which Oracle has said has begun. Oracle explained the sequence in its response to this report’s factual review, on October 12, 2026. The applications reach customers with Release 26D in phases, cohort by cohort, over about three months, and some customers receive them early. Public documentation follows each cohort, because an application in early access may still change on early customer feedback. No customer is named in the release, the deck, the documentation, or any of the 27 outlets Greyhound Research reviewed.

The count

The portfolio count is harder to pin down than the release suggests, and the difficulty predates Fusion Claw. Oracle announced 22 agentic applications on March 24, 2026, described 25 as “new” across three releases on April 9, with 24 distinct names among them, four more on June 29, and a further set on August 11 with no count stated. Its release of July 14 still spoke of “the 22 new Fusion Agentic Applications that were launched earlier this year,” and its product page, last updated August 4, 2026, lists 22. The September 29 release says 75.

The deck of September 25 lists 28 names under a headline of 25, and none of the four applications named in the release appears on that slide under those names. Two of the four, Ledger and Account Territory Growth Plan, appear on the Help Center page. Workforce Staffing and Shipping Consolidation appear under those names in no Oracle document reviewed other than the release. None of this is unusual for a portfolio that has grown by announcement for six months. It does mean that a buyer asking “which of the 25 can I switch on today” will not find the answer in a document yet.

Table 2 sets out what a buyer can check today, each row with its Oracle source and date. Table 3 sets out what a buyer cannot check yet, each row with where the evidence would appear if it existed. The second table is not a list of things Oracle lacks. It is a list of things Oracle has not yet written down, bounded to the documents reviewed and the date.

The two tables read together say something precise. Oracle has documented the switch, the meter, the stop, the policy compiler, the identity ceiling, the retention window, and the geography of inference for one model provider. It has not yet documented the runtime those things surround. Exhibit 2 draws the same distinction as a map, placing each element of the announcement in one of four states, documented, demonstrated, announced, or not addressed, so that a reader can see at a glance that the mechanism sits entirely in the second.

↑ Back to the section guide

The Machinery Beneath The Name: A Runtime Described, Not Yet Written

Oracle’s record holds two statements about runtimes, two months apart, and Greyhound Research reads them together. Oracle’s release of July 14, 2026, introducing the no-code and pro-code builder experience for agentic applications, said they run “inside Oracle Fusion Applications without separate runtimes, external orchestration layers, or custom infrastructure.” The release of September 29 says Fusion Claw provides “a new execution runtime that supports larger scale and longer-running work, continuous replanning, deterministic execution, and increased governed autonomy.”

The two are consistent. The July substrate, a supervisor agent orchestrating a graph of model, code, policy and approval nodes inside Fusion, was credited by Greyhound Research in July as “the machinery of a runtime rather than the language of a demonstration.” September adds a second runtime beneath the first, inside Fusion rather than beside it, and the second is the one that is not yet written down.

What Oracle has said about it

Oracle’s account of the second runtime is consistent and specific. At the briefing it said “we are creating a container dynamically,” with “an orchestration agent like Codex” provisioned to solve the task, after which “it dissolves.” The container does not, in Oracle’s words, have open access to the internet, it operates “only in the context of a fusion agentic application,” and acts “only through the agents that it’s supposed to operate with, which already have role-based access.”

Oracle has repeated the account in interviews. The container “comes into being only when the task is given” and “goes away when it doesn’t need to exist anymore once the task is complete.” “What distinguishes Fusion Claw from other approaches is that we take an orchestration agent and place it inside an isolated runtime container,” and “Claw remains inside Fusion on Oracle Cloud Infrastructure, adding an isolated execution container without requiring a separate customer environment.” Oracle has also said that the isolated environment prevents the language model from directly updating Fusion business objects. Oracle’s contract of September 29 meters a Fusion Claw action as a distinct action type, which is the one place the runtime exists in an Oracle document.

What the documentation says

That account is coherent and it is the right design. A planning model that never touches a business object directly, a container that holds no standing credentials because it does not stand, and execution that runs through agents already bound by roles, is how a careful architect would build governed autonomy inside a system of record. Nothing in Oracle’s documentation reviewed as of October 6, 2026 describes any of it.

The three documentation and contract documents that name Fusion Claw describe a configuration surface, an authority sentence, and the existence of a receipt. The word “container” does not appear in them. Nor does “orchestration agent,” “isolation,” or any statement of where the runtime executes. In the Oracle documents reviewed, Codex appears only as a coding assistant developers may use at design time. The nearest public description of “an orchestration agent like Codex” is OpenAI’s own, which on September 10, 2026 offered to “build and run cloud agents with the Codex harness, fully managed by OpenAI.” Oracle has not said whether its orchestration agent is Oracle’s own, which its statement that Fusion Claw was built from scratch on Oracle’s architecture suggests, or runs on a managed harness. The answer decides where the agent executes and under whose terms. Oracle has since said, in its response to this report’s factual review, that the deterministic execution runs where Fusion is hosted. Where the orchestration agent itself runs is not yet stated.

One constraint in the public record has been explained but not yet updated. The README of Oracle’s own AI Agent Studio repository states a “strict 60-second response limit” on Studio workflows. Asked at the briefing whether Fusion Claw removes it, Oracle said “that particular one doesn’t apply in this case.” Oracle has since explained why, in its response to this report’s factual review. Fusion Claw runs asynchronously and can take as long as a task needs, and the 60-second timeout is an internal one within AI Agent Studio, between the Oracle Cloud Infrastructure layer and a modular microservices architecture. The README, read on October 6, 2026, still carries the line on both its main and release-26C branches, and its changelog ends on August 21, 2026. No exception for autonomous outcomes is yet written in the README or in the 26C or 26D guides, so a buyer reading the README will find the limit and not the explanation. What was an open question about the product is now a lag in its documentation.

Two pieces with partial documentation

The policy compiler is real. Oracle’s guide and its Fusion Center of Excellence blog of August 5, 2026 document policy models that convert unstructured business policy into a deterministic decision function, with test cases that must pass and a publication gate before use. At the briefing, a California staffing policy’s rule that a plan below 95 percent coverage “must not be applied automatically” became code that, in Oracle’s words, “does not need to reinterpret this 95% number every single time.”

The learning loop is half real. Release 26D documents an AI Guidance store that can “capture explicit, programmatic learning from an AI agent or workflow and store it,” read under the user’s security context, with no deletion, retention or link to Fusion Claw documented on the two 26D pages that describe it. Whether that store is where the briefing’s “execution paths and outcomes” live, and what governs their deletion, is not written.

↑ Back to the section guide

What Governs The Outcome: Four Verbs Against Three Constructs

The release maps so neatly onto the four-verb test that the mapping deserves to be checked rather than admired. Greyhound Research applies the verbs to Oracle’s three constructs cell by cell. The envelope answers bound the authority. The harness answers name the identity, in words. The receipt answers read the log, in words. Table 4 records, for each cell, whether Oracle has documented it, claimed it, or neither, as of the documents reviewed on October 5 and 6, 2026.

Bound the authority

This is where Oracle is strongest, as it was in July, and the documentation has thickened since. Policy models generate code from written policy, test it, and require publication before a workflow can select it. Tool-level approvals exist. Release 26D adds “autonomy settings for the application.” The 26D feature text states the ceiling in a sentence any audit committee can read: “Fusion Claw operates within the organization’s established policies, permissions, approvals, risk thresholds, and authority limits. It executes actions within delegated authority, requests approval when required, and escalates exceptions when necessary. Although AI can plan and recommend actions, it can’t grant itself authority to act.”

Oracle has clarified, in its response to this report’s factual review, that the envelope “is not a separate artifact or a document.” It is the umbrella under which every agentic application operates within specified boundaries: role-based access controls, explicitly granted permissions, approvals, and the policies and procedures the business lays out. Fusion Claw executes inside that fence, within delegated authority, and by design does not go outside it. That places the envelope precisely, in controls Fusion already holds, and it moves the change-control question with it. What is not documented is change control over those controls when they govern an autonomous outcome. No Oracle document reviewed as of October 6, 2026 says whether a change to them is versioned, or whether widening the authority they delegate requires a second signature. The only change-control sentence in the record is the Fusion Center of Excellence’s advice that “a material change to rules, prompts, schemas, or templates deserves a new test run before publication.”

Name the identity

This verb is answered in words and borrowed in fact. The release says the harness applies controls “including identity” to each run. Oracle’s documentation of July 27, 2026 says what that identity is: “Every action carries the running user’s Fusion token,” and “During execution, the agent still cannot exceed the running user’s native Fusion data access.” Oracle was direct about it at the briefing: “what we are giving them is roles today,” with per-agent identity, “agent 123 took the action,” the next step, and “it’s your fleet, so you’re responsible for it.” Oracle has added, in its response to this report’s factual review, that Fusion Claw operates only under the role-based access given to the agents within the agentic application, and that any action outside the authority granted is subject to human approval.

Borrowed identity is bounded, and the ceiling is documented and conservative. It is not the test July set, which asked for an agent that runs “as its own principal, holding the least privilege its task allows, never borrowing a human’s credentials.” Buyers are ahead of the vendor here. UK Finance, in November 2025, supported the development of “know your agent (KYA) protocols” that “help verify agent identity, intent, and authorisation” when an agent acts for a consumer, and a US federal security guide for robotic process automation gave each unattended bot a distinct directory account and a named process owner as long ago as 2023. At least one vendor in the review set now documents a per-agent principal with a human sponsor and an approval cycle to widen its access, so the category’s gap on this verb has closed in one place.

Read the log

This verb is answered by the receipt, and the receipt as demonstrated is the richest per-run record Greyhound Research has seen described by any system-of-record vendor in the review set. On September 25 it showed the outcome, the owner (the logged-in user), the executor (the application), the delegated-authority line, a timestamp, the changes applied, the steps completed, an execution summary, the sources used with “time, date, stamped versions,” and the model used, and it could be exported and printed. Oracle has said that the receipt contains more context than a conventional audit log, including the full text of the policies used during the task.

None of that is yet in a public Oracle document. The 26D text says only that “Each outcome run provides an auditable, traceable outcome receipt.” No field list, export format, retention period, integrity control, or linkage to Fusion’s audit tables appears in any Oracle document reviewed as of October 6, 2026. Oracle has since listed what a receipt records, in its response to this report’s factual review, in eleven categories: execution duration, task understanding (the interpreted goal, scope, constraints and expected outputs), mode and context (execution mode, transaction restrictions and application context), the sources and capabilities used, capability metrics, inputs, outputs, execution steps, validation and exceptions, business metrics (baseline and recommended results), and generated artifacts. That is a fuller record than the ten fields shown on September 25. It does not yet state an export format, a retention period, an integrity control, or a link to Fusion’s audit tables. The retention profile option of 90 to 365 days is documented for “generative AI data” without saying whether receipts are covered.

The receipt also records something subtler than its missing fields. “Approval was not required. It operated within delegated authority” records the absence of a human. It does not, as shown, record the instrument that made the absence lawful: which policy, which version of it, who delegated the authority, and whether the owner named on the receipt could still have changed the result. Oracle’s list comes closest through the task’s constraints, the transaction restrictions and the validation performed. It does not name the policy version or the person who delegated the authority.

↑ Back to the section guide

The Verb Nobody Answers: Reversal Stays With The System Of Record

The fourth verb is the one the announcement does not mention, and Oracle’s own answer to it has not changed since July. Greyhound Research presses hardest here, because it is the verb a controller will ask about first. The July test was explicit about what reversal means: “A bad chain of decisions can be unwound through compensating transactions, with a named owner for the trigger.” It asked Oracle “what mechanism exists, and who triggers it?”

Oracle’s answer, given in July and repeated at the briefing, is that when an approved action turns out to be wrong the system falls back on the existing reversal mechanisms in the system of record, that the platform can show what happened and who approved it, and that it does not undo the transaction itself. At the briefing the same point came in plainer words. “Some are one-way doors, and some are kind of two-way revolving doors,” and “some things, yes, once it’s done, it is done,” which Oracle likened to a human making a mistake.

Why the answer deserves to be taken seriously

That answer should be weighed rather than scored against. Fusion’s general ledger reverses journals and its supply chain cancels shipments, as a system of record has always done, and each of those is a compensating transaction with a named human trigger, which is what the July verb literally asked for. An agent that could unwind its own postings would be an agent with more authority, not less. Oracle’s position is that the system of record already owns reversal, and for a single wrong posting that is true. Oracle restated the position in its response to this report’s factual review: transaction correction exists within Fusion Applications and is handled there.

What the constructs add

The Claw constructs leave reversal where Oracle says it belongs, in the system of record, and that is the finding. In the Oracle documents reviewed as of October 6, 2026, no document links an Outcome Receipt to a reversing entry. No document describes unwinding a chain of postings as a unit, which is what a full-auto outcome that reallocates fifty shifts or consolidates thirty shipments produces. No document describes a cancel for a single outcome in flight, or the state a partially executed outcome is left in when stopped. The one-way doors Oracle described, a shipment that has left or a payment that has cleared, are reached by no receipt. The receipt shows what was done. It does not show how to undo it.

What has arrived is the halt

Release 26D documents, as of September 21, 2026, a total budget and budgets at application and agent level, warning thresholds, and the option to “enable auto-stop to prevent new runs when a selected budget is exhausted.” It also documents the ability to “stop or resume all items when immediate operational control is required, including options to stop only new sessions or also stop all running sessions.” That is a kill switch, and it is an honest one: it stops spend and it stops the fleet. In June 2026 a named implementation partner had written that the only way to stop consumption was to unpublish agents, so this is a documented before and after in Oracle’s favor. It is also fleet-level. Stopping every running session to halt one outcome is a fire alarm, not a circuit breaker.

The gap belongs to the category, and the category has moved only at the halt level. Greyhound Research reviewed the public material of SAP, Salesforce, Microsoft, Workday, ServiceNow, Google Cloud, Amazon Web Services, UiPath, OpenAI and Anthropic between September 18 and October 5, 2026. Across it, vendors document cancel, pause and resume of running processes, credential revocation, per-call deny hooks, and, in two cases, a customer-facing export of the per-run record. None documents an unwind of posted work, and none records a reversal in a per-run record. The regulated industry itself has told its regulator that the switch may not be as simple as it sounds: in its response to the Monetary Authority of Singapore, the Asia Securities Industry and Financial Markets Association said a true kill switch “may not be feasible for some architectures” and asked for graded alternatives.

↑ Back to the section guide

Where The Reasoning Runs: Governed Autonomy Has A Geography

Governed autonomy has a geography, and the announcement does not mention it, so Greyhound Research maps where the frontier model runs. Every Claw outcome “begins with a frontier model,” and the frontier models the contract names at launch are OpenAI’s GPT 5.5 and Google’s Gemini 3.5 Flash. Where those models run is not where the customer’s Fusion environment runs.

Oracle documents this candidly for one provider. Its Help Center FAQ on OpenAI endpoint locations, with a build stamp of September 22, 2026, routes calls from Fusion environments in Mumbai, Hyderabad and Singapore to “Europe (EEA + Switzerland),” and calls from Sydney, Melbourne, Tokyo and Osaka to “USA.” Frankfurt and London route to Europe. Ashburn and Phoenix route to the United States.

What Oracle’s own infrastructure covers

Oracle’s infrastructure does not yet close that distance for every region. Its regional availability page for Oracle Cloud Infrastructure (OCI) Generative AI, updated September 24, 2026, lists Hyderabad, Frankfurt, London and Ashburn among the regions that host models, and does not list Mumbai, Singapore, Sydney or Melbourne. At the briefing, asked how a customer in Sydney or Mumbai should think about sovereignty when the reasoning leaves the country, Oracle did not promise in-region inference. It is “constrained by where OpenAI models or Gemini models are available,” GPT-5.6 Luna “will be available in US and maybe in Frankfurt at best,” and running Oracle’s own models in restricted realms is “definitely on our radar” but “not there fully yet.”

The contract adds one footnote a government buyer should read first: “Certain commercial third-party models, including commercial OpenAI models, are not supported in U.S. Government Cloud Services.”

What is not yet written

Oracle has now answered the first half of the question, in its response to this report’s factual review. A Claw outcome’s reasoning runs on the frontier model, in the locations where that model’s provider, OpenAI or Google, makes it available, and the deterministic execution runs where Fusion is hosted. For OpenAI models, the FAQ’s routing table therefore applies to a Claw outcome’s planning. The FAQ covers OpenAI only, and no Oracle document reviewed describes a customer control that pins inference to a region or to an Oracle-hosted model. Oracle has said that where a third-party model is used it relies on zero-data-retention agreements under which the provider cannot store or train on the prompt. That is a contractual control on retention. It is not a control on location. In the week after the announcement, no outlet Greyhound Research reviewed raised the question, and no buyer raised it in any venue Greyhound Research reached.

↑ Back to the section guide

The Rules Buyers Live Under: A Four-Part Test The Receipt Half Meets

An institution that delegates authority to software is still the institution the regulator will write to. Greyhound Research therefore reads the regulators’ texts in seven jurisdictions against the receipt. The texts differ in date and in force. The rule they converge on has four parts. The institution answers for the agent. A competent person can halt it before its output takes effect. The record lets a third party reconstruct what it did. And an affected person can reach a human who is empowered to change the result. No text in force in any of the seven requires an institution to reverse a posted autonomous action, which is why the fourth verb is a buyer’s problem rather than a compliance finding.

Europe, where the text is most specific

Article 14 of the EU AI Act requires that the humans overseeing a high-risk system be able “to otherwise disregard, override or reverse the output of the high-risk AI system” and “to intervene in the operation of the high-risk AI system or interrupt the system through a ‘stop’ button or a similar procedure that allows the system to come to a halt in a safe state.” Article 12 requires a high-risk system to record events automatically over its lifetime. Article 26 requires deployers to assign oversight to people with the competence and authority to exercise it, to keep the logs the system generates for at least six months, and to inform workers’ representatives before putting a high-risk system to use at a workplace.

Those obligations apply to Annex III systems from December 2, 2027, and Annex III point 4(b) covers systems used “to allocate tasks based on individual behaviour or personal traits or characteristics.” A Workforce Staffing application that allocates shifts is within that wording by description. In the United Kingdom, the automated decision-making provisions of the Data (Use and Access) Act have applied since February 5, 2026, with safeguards that let a person contest a significant automated decision and obtain human intervention. The Mills Review for the Financial Conduct Authority, published in July 2026, said what the oversight has to look like: “It will not be enough to say that a person remains ‘in the loop.’”

The United States, where the text points back at the institution

The United States has written the opposite kind of text and reached the same place. The federal banking agencies’ model risk guidance of April 17, 2026 states in a footnote that “Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance,” and then returns the problem to the bank: its own governance “should guide the determination of appropriate governance and controls” for anything the guidance does not cover. Colorado’s law, from January 1, 2027, requires meaningful human review, to the extent commercially reasonable, by someone with authority to approve, modify or override a consequential automated decision. California’s regulations from the same date exempt only decisions over which a human reviewer has authority of the kind the approved text defines.

India, Australia, New Zealand and Singapore

India’s central bank has asked the question three times in four weeks. A deputy governor of the Reserve Bank of India put it on September 11, 2026 as a sentence Oracle’s envelope could be built on: “Technology can distribute a service, but it cannot distribute responsibility.” A second deputy governor said on September 24 that the use of AI must be accompanied by human oversight and clear accountability. The governor said on October 3 that the bank’s model risk guidance, with human oversight and explainability at its center, remains in draft. India’s market regulator has said that its forthcoming framework would require kill switches and humans in the loop. Its data protection rules, notified in November 2025, bring their full obligations into force eighteen months after notification.

Australia’s transparency obligation for decisions made solely by computer programs applies from December 10, 2026. New Zealand has no AI-specific provision, and its privacy principles, one of them in force since May 1, 2026, require reasonable steps to ensure information is accurate before it is used. Singapore’s model governance framework for agentic AI asks deployers to be able to “stop agent workflow and escalate to a human supervisor” and states that “An agent should have its own unique identity.” Its cyber security agency advises human approval “for high-stakes or irreversible actions such as financial transactions, executing code in production, deleting critical data or sending external communications.”

The auditors

The audit profession is writing its text now. The International Auditing and Assurance Standards Board’s proposed revision of ISA 500, open for comment until December 15, 2026, adds an “Authenticity” attribute for audit evidence: that the information “has been generated by or provided by a source authorized to do so” and “has not been altered in a manner that makes it inadequate for the audit.” It also asks auditors to test the controls over information an entity’s own systems produce. A receipt that cannot be exported, retained on the auditor’s terms, and shown to be unaltered will not meet that attribute, whatever it displays on screen. Exhibit 3 sets the receipt as demonstrated against the four parts of the converging rule, field by field, and marks what the demonstration did not show as exactly that, never as absent from the product.

The clause that falls on the headline application

The rules have one more clause that neither the announcement nor any coverage of it mentions. In the Netherlands, guidance for works councils published in December 2025 treats an AI system that automatically plans rosters or working hours as a matter requiring the works council’s consent under Article 27 of the Works Councils Act. The European Parliament’s resolution of December 17, 2025 asks that decisions on pay, dismissal and discipline “always be taken by a human being.” The European Commission, after a consultation on algorithmic management that closed on September 28, 2026, has said a proposal “will be presented later this year.” For rostering in the Netherlands the envelope is a consent matter before it is the controller’s document. The consent, where it exists, comes before the dial.

↑ Back to the section guide

The Name Oracle Chose: A Convention Joined, A Mark Unresolved

Two questions about the name belong in a buyer’s security and contract review, and Greyhound Research gives them one bounded section rather than a footnote. Neither is about the word itself. Both are about what sits under it, and whether it will still be called that in a year. Oracle has since answered the first of them in writing.

Provenance

Oracle has said in one interview that Fusion Claw’s separation of reasoning and execution “was all inspired by OpenClaw and where OpenClaw started,” and in another that it was built from scratch with Oracle’s architecture and proprietary intellectual property. The difference between inspired and built-on is the difference between a design idea and a dependency, and it is the one a chief information security officer will ask about first. Oracle has now settled it in its response to this report’s factual review: there is no connection to OpenClaw, Fusion Claw is Oracle’s own architecture, and it does not depend on OpenClaw runtime components. The interview line describes a lineage of ideas at most, and the dependency question is closed.

Oracle joined a naming convention rather than inventing a risk. Cisco, Automation Anywhere, NVIDIA and Microsoft had each put the word on a product between March and June 2026. No security researcher, foundation or community platform Greyhound Research could reach has objected to Oracle’s use of it. The provenance question was therefore never a reputational one. It was a dependency question, and Oracle has now answered it.

The mark

“Fusion Claw” is a product of Wilcox Industries, a maker of headborne night vision, communications and targeting equipment. Wilcox holds a live United States trademark application for FUSION CLAW in Class 009, published on February 17, 2026 and allowed on April 14, 2026, with a statement of use due on or about October 14, 2026. No Oracle application for the mark, and no Oracle statement on it, was found as of October 6, 2026. For a buyer the consequence is practical rather than legal: a product name that may change is a name to keep out of the policy documents, the delegation schedules and the contract exhibits that will outlive it.

↑ Back to the section guide

What CIOs Are Weighing: Accountability Before Capability

The buyer side of this record is quieter than the regulators and points the same way. Greyhound Research reads it here from the public record of named buyers, industry bodies and classified surveys, and from Greyhound CIO Pulse 2026 and Greyhound Fieldnotes (advisory engagements with buy-side clients), to set out what technology leaders are weighing when software is offered the authority to act. None of it is about Fusion Claw by name. As of October 6, 2026, no Oracle customer and no named CIO, CFO, CHRO or audit chair has spoken about the product in any venue Greyhound Research could reach. All of it is about the question the product raises.

The question buyers ask first is not capability

In the external record, accountability is the gating question and model quality is raised by none of the buyers, bodies or surveys found. Greyhound CIO Pulse 2026, as published in Greyhound Research’s research report of July 20, 2026, across more than 1,000 technology leaders, finds fewer than one in three can point to measurable outcomes across most of their funded AI portfolio. A leader who cannot yet show the outcome of an assistant is being asked to answer for the outcome of an executor, and that is the order in which the questions arrive.

The named voices say it without a survey. An assistant commissioner of the Australian Taxation Office has said that agentic components stay switched off because “in the end, we’re accountable for the outcomes.” Dell’s chief financial officer has said that agents now do reconciliations and journal entries and that “The accountability level is still there.” One has kept the dial closed and one has opened it, and both describe the same thing: the signature did not move.

The record splits by function

Greyhound Research’s reading of the buyer record is that it splits by function, and that the split should be stated rather than averaged. Finance and HR are cautious. The named finance and HR voices in the record speak of accountability before capability, of agents doing the reconciliations with the signature unmoved, and of agentic components kept switched off. A poll of more than 3,400 digital trust professionals by their professional body in March 2026 found 56 percent do not know how fast they could halt an AI system and 20 percent do not know who would be accountable if one caused harm, which is the governance function saying it has not yet set the envelope.

Operations is further along. In Greyhound Research’s reading, the agents already running without a person in the loop sit in IT operations, supply chain and service, where a wrong action is usually reversible and visible, and the dial has moved there in many estates. The split matters for Oracle’s four applications, because two of them, Ledger and Workforce Staffing, sit in the functions that are most cautious, and the other two sit in operations, where the question is not whether to delegate but how much.

What buyers ask for, and what they do not

Where buyers ask for a control, they ask for halt and suspend, not for unwind. The 56 percent figure is about stopping, the Singapore framework is about stopping, and the regulated industry’s own submission to its regulator is about whether stopping is even feasible for some architectures. Nobody in the buyer record has asked a vendor to reverse a posted outcome, which is consistent with the finding in The Verb Nobody Answers: reversal has always lived in the system of record, and buyers know where it lives.

Where buyers lead the vendor, it is on identity. The case for an agent carrying its own verifiable identity and authorization was made on the buyer side, in UK Finance’s submission of November 2025, before Oracle documented one. Works councils in the Netherlands had guidance on automated rostering before the product that automates rostering was announced. The buyer record is ahead of the vendor record on the two verbs that are about people, who acted and who agreed, and behind it on the two that are about machinery.

The mental model that travels

The simplest way for a technology leader to hold all of this is the one the enterprise already uses for a contractor. Oracle is offering a very fast contractor. The envelope is the contract, the harness is the building pass, and the receipt is the paperwork. A buyer who would not let a contractor on site without a signed scope, a pass that opens only certain doors for one shift, and an expense report at the end, now has a vendor that has built all three into the software. What it does not yet have is a scope written on the buyer’s side.

↑ Back to the section guide

What An Outcome Costs: A Meter That Moved On Announcement Day

The economics claim in this announcement is a claim in AI Units, and the rules of the meter changed on the day the announcement was made. Greyhound Research reads the meter before the marketing. “Zero Oracle cost is not zero customer cost,” Greyhound Research wrote in July, before it had read Oracle’s meter. The earliest version of Oracle’s Fusion Cloud Service Descriptions in the record, effective September 10, 2026, carried one. On September 29 the same document was replaced at the same address with a new version, effective that day. The replacement moved the included allowance and the effective date of the pricing table, renamed one model tier, and added a Frontier tier with its own factors and a Fusion Claw row, with a single footnote to mark one of the changes.

The claim, in Oracle’s words

The release says the separation of reasoning from execution “provides Fusion Applications customers with a potential economic advantage by applying AI reasoning only where intelligence is needed and using efficient deterministic computation for high-volume execution.” The deck of September 25 marks “AI units used” against planning and “AI units not used” against deterministic execution. Oracle has said in interviews that customers consume units when a model is planning and that, once a plan is developed and the platform executes it, no units are consumed, and that reasoning costs can be cut by “50%-plus” where an approved plan is reused. It has offered no benchmark, and has said that frontier models “need capacity” and “they’re fairly expensive.”

The meter, in the contract’s words

Usage is measured in AI Units, calculated as the input and output tokens of the model in use divided by 10,000, rounded up, and multiplied by an Action Value Factor. Each Fusion production environment now includes “a total of 30,000 AI Units” a month, up from 20,000 in the September 10 version. The factors depend on the model tier and the action type, and the table that sets them carries a row for Fusion Claw, read by a human on October 6, 2026.

A Fusion Claw action, whose typical uses are deep research, autonomous execution and planning, is metered at 20x on a Frontier model and at 10x on a model the customer brings. It is not available on the Basic model, and it is not available on a Balanced model either. Balanced is the tier Oracle renamed from Premium on September 29, and its default model, GPT-5.6 Luna, is the one Oracle’s own guide marks as the Studio default. A Claw outcome runs on the most expensive tier Oracle meters or on the customer’s own model, and on nothing in between.

The arithmetic a buyer can do

The arithmetic is short, and it stops where Oracle’s disclosure stops. Oracle’s Fusion Insider blog of August 27, 2026 states that “an AI Unit is worth one cent,” and Oracle’s Global Price List of September 10, 2026 confirms it, at USD 1,000 for 100,000 pooled units. At that value, a Fusion Claw action on a Frontier model costs twenty cents for every 10,000 tokens of combined input and output. The 30,000 units included with each production environment buy 15 million such tokens a month before any other agent, document or approval draws on the same pool. The Fusion Agentic Applications subscription, listed at USD 500,000 a year, carries 2,500,000 pooled units a month.

What no Oracle document states is how many tokens a Claw outcome consumes in planning and re-planning, which is the only number that turns a factor into a cost. Units per outcome is a defined gap in the record, not an estimate this report will supply.

Two definitions, and Oracle’s answer

The blog says units “are deducted only for AI actions delivered and completed, not for intermediate steps like API calls.” The contract lists “Reasoning steps” and “Approve transactions” among the typical uses that are metered under the General action type, at 20x on a Frontier model and at 0x on the Basic model. Oracle has answered the question this raised, in its response to this report’s factual review. A frontier model consumes units when it understands the objective, reasons and plans. Deterministic execution, the computation that runs under the plan, “does not involve any metered AI actions.” Where a workflow agent inside the application performs a simple action, it can use the Basic model hosted on Oracle Cloud Infrastructure, which the contract meters at 0x. The contract adds one condition a budget should note. It describes the Basic model as designed for basic use cases “at modest volumes,” and recommends a non-Basic model for agentic use cases involving planning and tool calling. The tier an agent step runs on therefore decides whether it draws units. Two named implementation partners and one independent practitioner, writing in June 2026, had already documented what the included units do and do not do: they reset monthly, they are consumed in development and test as well as production, and document generation is never free. Table 5 sets the two contract versions beside each other and states what each line means for a Claw outcome.

↑ Back to the section guide

The Autonomy Proof Test: Documented, Claimed, And Owed

With the announcement a week old and the newest documentation four days old, Greyhound Research proposes the evaluation decision as a proof test rather than a verdict. The test has three columns. What is documented, with a date and a document. What is claimed, by whom and where. And what is owed, meaning what Oracle has said or implied that a buyer cannot yet verify, with the place the proof would appear. Exhibit 4 fills them.

The decision, and where to start

The decision a buyer faces is not whether to switch Fusion Claw on. It is where to set the dial, per class of outcome, and who signs for the setting. Oracle’s product gives the customer the dial and says nothing about where it should start. The regulators’ texts, the buyer record and Oracle’s own statements all point the same way. Begin where the plan is staged for review. Move to full auto one outcome class at a time, when the receipt for that class has been read by the person who would have to answer for it. Oracle expects organizations to start with more human review. The implementation partner that has written the clearest first step says the same thing: run one high-volume process in review mode, compare the cost per outcome against the baseline, and extend autonomy as the results support it.

Three conditions that follow from the evidence

The first is that the dial belongs to the owner of the outcome, not the administrator of the application, because the receipt names the owner and the regulators name the institution, and neither names the person who found the switch. The second is that an outcome class is eligible for full auto only when its postings can already be reversed in bulk by the system of record and the controller knows how, because nothing in the Claw constructs will do it for them. The third is that the unit budget is set before the dial moves and set per application, which is the object Release 26D provides, with the knowledge that the factors beneath it can change on thirty days’ notice.

↑ Back to the section guide

The Greyhound CXO Runbook: Five Actions Before Any Dial Moves

Five actions follow from the evidence, not from a template, and Greyhound Research sets them out in the order they depend on one another: each has a chair, and the first is the one the other four rest on. None of them waits for Oracle, because the authority a Claw outcome runs under is set in controls the buyer already holds.

1/ Set the envelope and the unit cap before any dial moves. The dependency is the delegation of authority matrix. Release 26D documents a switch per application that enables autonomous outcomes, and Oracle’s own feature text says the AI “can’t grant itself authority to act.” The authority therefore has to come from the customer’s own controls, which Oracle confirms are what the envelope consists of: roles, granted permissions, approvals and policies. The documented objects that hold it today are policy models with test cases and a publication gate, tool-level approvals, and the application’s autonomy settings.

Before any application is enabled, write down the outcome classes it may pursue, the threshold on each, the policy and version that governs it, the name of the person who delegates the authority, and the second signature required to widen any of it. Then set the application’s AI Unit budget in the 26D budget controls with auto-stop on, because that is the object Oracle provides.

In a European workforce the envelope for rostering is a negotiated matter: works council consent in the Netherlands precedes the dial today, and the duty to inform workers’ representatives reaches every EU employer deploying a high-risk system on December 2, 2027. The decision sits with the owner of the outcome, the controller for Ledger, the head of HR for Workforce Staffing, the supply chain owner for Shipping Consolidation, countersigned by the chief information officer, and never the other way round.

2/ Put the region of reasoning and the model in writing before regulated data enters any mode. Research mode plans with a frontier model exactly as full auto does, so residency is a precondition of the first run, not of the last.

Oracle’s FAQ routes OpenAI calls from Mumbai, Hyderabad and Singapore to Europe and from Sydney and Melbourne to the United States, its OCI regions page lists Hyderabad but not Mumbai, Singapore, Sydney or Melbourne as hosting models, and its contract excludes commercial OpenAI models from US Government Cloud. Oracle has said, in its response to this report’s factual review, that a Claw outcome’s reasoning runs where the model’s provider makes the model available, so for OpenAI models the FAQ’s routing applies. The same contract puts every Claw action on a Frontier model or on a model the customer brings.

What the buyer needs is that answer as a written term in the buyer’s own contract, stating where a Claw outcome’s planning executes for the customer’s region, which model is pinned to each outcome class, whether a model change is recorded on the receipt with its version, and whether an Oracle-hosted model in a restricted realm is available for the outcome classes that need one.

A buyer who cannot get the term should treat the region of reasoning as unknown and keep regulated data in work assistance until it is known. The decision sits with the chief information security officer, with the data protection officer holding the term and procurement holding the signature.

3/ Classify every outcome class by its door before choosing the first full-auto class. Oracle said it at the briefing: some are one-way doors and some are revolving doors. Nothing in the Claw constructs reverses anything, and Oracle’s July answer stands, so reversal is the system of record’s own compensating transaction with a human trigger, one posting at a time.

The instrument is a register, held by the process owner and read by internal audit, that marks each outcome class as reversible in bulk by a rehearsed procedure, reversible one entry at a time, or not reversible once executed. It bars the third class from full auto and keeps the second in research mode until the first class has run clean for a period the audit committee sets.

A payment that has cleared and a shipment that has left are reached by no receipt, and the register is where that fact is written down before the dial is found. The decision sits with the process owner and the head of internal audit together.

4/ Run one class in research mode, export the receipt, and read it against the audit tables. The receipt is the best object in the announcement and the least documented. It showed ten fields on September 25, and no Oracle document reviewed lists them or states an export format, a retention period, an integrity control, or a linkage to Fusion’s audit tables. Oracle has since listed eleven field categories in writing, and the export, retention, integrity and audit-table questions remain.

Run one outcome class in research mode in a non-production environment, where units are consumed too, and export every receipt. Have the person who would sign for the outcome read it beside the audit table entries for the same postings, checking three things: that the identity on each posting is one the audit table can attribute, that the line “Approval was not required” can be traced to a named policy, its version, and the person who delegated it, and that the receipt survives in a form the auditor can retain on the auditor’s terms.

The authenticity attribute the International Auditing and Assurance Standards Board has proposed for audit evidence, open for comment until December 15, 2026, is the test the external auditor is likely to apply once the revised standard is adopted. A receipt that fails it in a test environment will fail it in production. The decision sits with the audit committee chair, through the controller.

5/ Measure units per outcome from those runs, then set the budget from the measurement. The factor is known and the tokens are not. A Claw action on a Frontier model is 20 units per 10,000 tokens, at the cash value Table 5 records, and “Approve transactions” is metered under the General action type, at the same factor on a Frontier model and at 0x on the Basic model. Oracle has stated in writing that deterministic execution draws no units, and the research-mode runs are where a buyer sees that on the meter.

Take the research-mode runs of the fourth action, read the consumption that the 26D budget controls record for the application, and divide by outcomes. Set the application’s budget and warning threshold from that figure rather than from “a potential economic advantage.” Three things to hold in mind while doing so: the included allowance is shared with every other agent in the environment and buys the tokens Table 5 records, the factors can change on thirty days’ notice, and the included allowance changed and a tier and a row were added once already, on announcement day.

The decision sits with the chief financial officer, and the number belongs in the business case before the business case is signed.

↑ Back to the section guide

Questions Ahead For Oracle: What Only Oracle Can Settle

What a buyer cannot reach alone, Oracle can. Every gap recorded in this report is a gap in documentation or a limit of a demonstration, and each can be closed by Oracle, most of them by writing down what it has already said in public. Greyhound Research puts the ten questions below in the order the evidence raises them, and will publish Oracle’s answers as given. Oracle answered parts of seven of them in its response to this report’s factual review, and those answers are printed in the sections above. What follows is what remains open.

1/ Which applications are Claw-powered today, by name, in which release and region, and can Oracle name one customer running any of them in any mode? The release says 25 are available today, the Help Center names fifteen, and the deck lists 28 under a headline of 25. Two of the four headline applications appear under their release names in no Oracle document reviewed other than the release, and Oracle’s statements on timing range from “available today” to the fourth quarter. No customer is named anywhere. Oracle has explained that the applications reach customers with Release 26D cohort by cohort, with documentation following each cohort. A list with release levels and dates, and one named early-access customer, would settle whether “available today” describes software or an intention. An answer will be printed as a documented row in Table 2, and the first owed item in Exhibit 4 will change columns.

2/ Where is the Fusion Claw runtime documented, by document identifier and page, and what is the orchestration agent? Oracle has described a container created per outcome and dissolved on completion, and an orchestration agent “like Codex,” and no Oracle document reviewed contains the word container. Oracle has said the deterministic execution runs where Fusion is hosted. Does the orchestration agent’s container run there too? Is the orchestration agent Oracle’s own or a managed harness from a model provider? Oracle has answered the duration question: an outcome runs asynchronously, and the README’s “strict 60-second response limit” does not bind it. Will the README and the guides say so? The answer settles whether the machinery is a product fact or an interview. An answer will move the runtime from the demonstrated lane of Exhibit 2 to the documented lane, cited by document and page.

3/ Since the Enterprise Operating Envelope is an umbrella over existing roles, permissions, approvals and policies, which of those controls are versioned when they govern an autonomous outcome, who may change them, and what is the default autonomy level of each capability when Release 26D is applied? The release names the envelope and the documentation names policy models, approvals, and “autonomy settings for the application.” Oracle has confirmed that controls of this kind make up the envelope. None of the Oracle documents reviewed says whether a change to them is versioned, or whether widening the delegated authority or raising the autonomy level requires a second approver. The default matters more than any of it: the first action in the Runbook assumes the dial starts closed, and Oracle has not said that it does. An answer will be printed in the envelope row of Table 4, and the Runbook’s first action will name the controls.

4/ Under which identity are a Claw outcome’s postings recorded in Fusion’s audit tables today, and on what date will a per-outcome principal be available? Oracle’s documentation of July 27, 2026 says every action carries the running user’s token, and Oracle said at the briefing that “what we are giving them is roles today” with per-agent identity as the next step. An auditor reading the audit table for a posting a Claw outcome made will see a name. Whose, and when will it be the outcome’s own? The answer settles what the audit table attributes today and the date from which it attributes an agent. An answer will change the identity cell of Table 4 from borrowed to owned, with the date beside it.

5/ What are the Outcome Receipt’s export format, retention period, and integrity control, and does the receipt record the policy identifier, its version, the threshold evaluated, and the person who delegated the authority? Ten fields were shown on September 25, and Oracle has since listed eleven field categories in writing, none of them yet in a public document. The retention profile option covers “generative AI data” without saying whether receipts are included. “Approval was not required. It operated within delegated authority” records that no human acted. The receipt an audit committee can sign for records which instrument made that lawful, and a specification would settle whether this one does. A specification will be printed field by field against Exhibit 3, with every “not seen” mark re-scored.

6/ Is there a cancel for a single outcome in flight, in what state is a partially executed outcome left when all sessions are stopped, and does any Claw construct link a receipt to Fusion’s reversal mechanisms for a chain of postings? Release 26D documents a budget auto-stop and a stop for all running sessions, both fleet-level. Oracle’s answer since July is that the system of record reverses and the platform does not, and it restated that answer in its response to this report’s factual review. A buyer also needs to know whether Oracle’s contract offers any warranty or indemnity for loss caused by an outcome executed within delegated authority, because if it does not, the envelope is the only instrument of recourse the buyer has. An answer will close the reversal row of Table 4 or confirm it as a category gap, and the warranty position will be quoted as given.

7/ For a customer in Mumbai, Hyderabad, Singapore, Sydney, or Melbourne, can a Claw outcome’s frontier-model reasoning be pinned to a region or to an Oracle-hosted model in a restricted realm, and where does Gemini 3.5 Flash run for each? The endpoint FAQ answers the routing for OpenAI chat and agent calls, and Oracle’s response to this report’s factual review extends it to autonomous outcomes: reasoning runs where the model’s provider makes the model available. At the briefing the answer for GPT-5.6 Luna was “US and maybe in Frankfurt at best.” A regulated buyer in any of those cities cannot write the second action of the Runbook without this answer, and Oracle is the only party that can give it. An answer will be printed as a routing statement beside the FAQ’s table, region by region.

8/ How many units did each of the four named applications consume per outcome in Oracle’s own demonstrations, which model tier do the agent steps around the plan run on, and was thirty days’ notice given for the factors effective October 1, 2026? Oracle has answered the first question this report asked here: deterministic execution draws no units, and a simple agent step on the Basic model is metered at 0x. The included allowance, the tier names, and the effective date changed on September 29, and a Frontier tier and a Fusion Claw row were added, with only the rename footnoted. A units-per-outcome range for Ledger, Workforce Staffing, Shipping Consolidation, and Account Territory Growth Plan would turn “a potential economic advantage” into a number a controller can test. A range will be printed in Table 5 against the units-per-outcome row, and the Runbook’s fifth action will cite it.

9/ For Workforce Staffing and the payroll and compensation applications in Europe, what does Oracle provide to help a deployer meet the duty to inform workers’ representatives from December 2, 2027, and the Dutch works council’s consent right today? Rostering by software is a consent matter under Article 27 of the Dutch Works Councils Act in guidance published in December 2025, and Annex III of the AI Act reaches task allocation “based on individual behaviour or personal traits or characteristics.” Nothing in Oracle’s release, deck, briefing, or documentation reviewed mentions a works council. A template, a data sheet, or a sentence would do, and its absence will be noticed first by the customers Oracle is most likely to headline. Whatever Oracle provides will be named in the Runbook’s first action, beside the works council consent.

10/ Has Oracle filed or cleared the mark FUSION CLAW in any jurisdiction? Oracle has answered the provenance question in writing: Fusion Claw has no connection to OpenClaw and does not depend on OpenClaw runtime components, and The Name Oracle Chose prints that position. An equipment maker holds a live application for the same mark in the same class with a statement of use due on or about October 14, 2026, and no Oracle filing or statement was found. One sentence from Oracle’s counsel would close it, and Oracle’s position will be printed in The Name Oracle Chose as given.

↑ Back to the section guide

The Four-Verb Verdict: The Final Greyhound Standpoint

Greyhound Research’s final judgment is that Oracle Fusion Claw is new machinery and not new vocabulary. As the evidence stands, with Oracle’s documents read on October 6, 2026 and its written answers of October 12 added, the machinery is described by Oracle, metered in a contract, and switched on by a configuration page, with the chapter that would let a buyer read it still unwritten. The idea beneath it is a more mature one than the category has generally offered: the delegation of authority matrix that every enterprise already holds, made executable, with a receipt as the evidence that software obeyed it. Oracle has built the part that obeys. The part that delegates is the buyer’s, and most of what the Runbook asks is setting that delegation down in controls Fusion already holds.

On the four verbs the ledger has moved, and the movement is uneven in exactly the way the evidence is. In July, on the public record, Oracle was credited on one verb. On bound the authority, the documentation has thickened since, with policy models that compile, test and gate, with autonomy settings per application, and with a sentence, “it can’t grant itself authority to act,” that belongs in every buyer’s envelope, so the credit stands and strengthens.

On read the log, the receipt as demonstrated is the richest per-run record Greyhound Research has seen described by any system-of-record vendor in the review set. Two vendors in that set document an export of their per-run record and Oracle does not yet, so the credit for it is conditional on a specification that does not exist. On name the identity, the answer is documented and it is borrowed: the running user’s token, bounded by that user’s access, with a per-outcome principal described as the next step.

On reverse the action, nothing in the Claw constructs is addressed to it, the system of record’s compensating transactions invoked by a named human remain the only mechanism, and no vendor in the review set has gone further. The gap belongs to the category, bounded to the documents reviewed and the date.

One verb documented, a second within reach of a document, a third bounded and borrowed, a fourth left with the system of record, by Oracle and by every vendor reviewed. Confidence rose on two and held on two.

What supports the judgment is dated and short. Three Oracle documents name the product, a contract meters it as a distinct action type, a halt exists at fleet level, the identity ceiling is written down, and the retention window and the geography of inference are disclosed for one provider. Oracle’s written answers of October 12, 2026 settle four points that were open: how long an outcome may run, what the deterministic layer consumes, where the reasoning runs, and the provenance of the runtime. The regulators’ texts in seven jurisdictions converge on a four-part rule that the design fits better than any receipt a journal entry has ever carried. What remains conditional is the list the proof test owes: the runtime chapter, the receipt specification with export and retention, versioning and a second signature on the controls that make up the envelope, a per-outcome cancel and the state of a stopped outcome, units per outcome, a control to pin the region of reasoning, the 25 by name, and a customer. Every one of those is a documentation finding, and Oracle can close most of them with text.

What would warrant reconsideration is equally specific, and it runs both ways. The judgment improves the day the guide chapter that names the container is published, the day a receipt specification with an export format and a retention term appears, and the day a named customer runs a full-auto outcome whose receipt has been read by the person who answers for it, which Release 26D makes possible in production from November 20, 2026. It worsens if the receipt cannot leave the screen, or if the factors beneath the meter move again without notice.

The dates to watch are Oracle AI World from October 25 to 28, 2026, the Wilcox statement of use due on or about October 14, 2026, the European Commission’s algorithmic management proposal promised for late 2026, and December 2, 2027, when the AI Act’s oversight and worker-information duties reach the application Oracle chose to headline.

For technology leaders the change is procedural before it is strategic. The product gives them a dial and a receipt. The institution still has to supply the signature, and the order of work is the Runbook’s: set the envelope, fix the region of reasoning, classify the doors, read one receipt against the audit table, and measure the units before the budget is set. Greyhound Research will revisit this position at each of the dates above, and no later than the first production cohort of Release 26D.

Oracle has put the receipt on the screen and the switch in the documentation. The signature the receipt records is still the buyer’s to give, and that is how it should be.

↑ Back to the section guide

Sources And References: Every Document Behind Every Claim

Every load-bearing claim in this research report traces to one of the sources below, presented in Harvard reference format and grouped by class. In-text author-date citations are omitted, as they serve academic writing rather than research intended for the executive desk, and in the web edition each source is hyperlinked in the sentence that carries it. Everything in the first four classes is Oracle’s own account of itself and proves what Oracle said rather than that the thing is true. Two sources are disclosed briefing material that is not public and are listed as such. The category review set named in the body (SAP, Salesforce, Microsoft, Workday, ServiceNow, Google Cloud, Amazon Web Services, UiPath, OpenAI, and Anthropic) was reviewed between September 18 and October 5, 2026, and the entries listed under that class are the ones that support a specific sentence. Standing pages that carry no publication date are marked n.d. with their visible update stamp. All sources were accessed on October 5 and 6, 2026, except where an entry states otherwise.

Oracle official announcements and corporate documents

Oracle Corporation (2026a) Oracle Introduces Fusion Agentic Applications. Press release, March 24. Available at: https://www.oracle.com/news/announcement/oracle-introduces-fusion-agentic-applications-2026-03-24/

Oracle Corporation (2026b) Oracle Introduces Fusion Agentic Applications for Finance and Supply Chain. Press release, April 9. Available at: https://www.oracle.com/news/announcement/oracle-introduces-fusion-agentic-applications-for-finance-and-supply-chain-2026-04-09/

Oracle Corporation (2026c) Oracle Introduces Fusion Agentic Applications for Customer Experience. Press release, April 9. Available at: https://www.oracle.com/news/announcement/oracle-introduces-fusion-agentic-applications-for-cx-2026-04-09/

Oracle Corporation (2026d) Oracle Introduces Fusion Agentic Applications for HR. Press release, April 9. Available at: https://www.oracle.com/news/announcement/oracle-introduces-fusion-agentic-applications-for-hr-2026-04-09/

Oracle Corporation (2026e) Oracle Adds New Fusion Agentic Applications to Help Customers Improve Supply Chain Performance. Press release, June 29. Available at: https://www.oracle.com/news/announcement/oracle-adds-new-fusion-agentic-applications-to-help-customers-improve-supply-chain-performance-2026-06-29/

Oracle Corporation (2026f) Oracle Introduces AI-Native Builder Experience to Create and Run Agentic Applications in Oracle Fusion Applications. Press release, July 14. Available at: https://www.oracle.com/news/announcement/oracle-introduces-ai-native-builder-experience-2026-07-14/

Oracle Corporation (2026g) Oracle Adds New Fusion Agentic Applications and AI Agents to Help Organizations Improve Talent Management. Press release, August 11. Available at: https://www.oracle.com/news/announcement/oracle-adds-new-fusion-agentic-applications-and-ai-agents-to-help-improve-talent-management-2026-08-11/

Oracle Corporation (2026h) Oracle Fusion Cloud Service Global Price List. Price list dated September 10, Oracle-hosted regional copy, marked “For educational purposes only. Subject to change without notice.” Available at: https://www.oracle.com/latam/a/ocom/docs/corporate/pricing/oracle-fusion-cloud-global-price-list.pdf

Oracle Corporation (2026i) Oracle Fusion Cloud Service Descriptions, v091026. Contract document, effective September 10, since replaced at the same address by v092926. Formerly available at: https://www.oracle.com/contracts/docs/oracle-fusion-cloud-service-desc-1843611.pdf

Oracle Corporation (2026j) Oracle Extends Fusion Agentic Applications with Introduction of Fusion Claw. Press release, September 29. Available at: https://www.oracle.com/news/announcement/oracle-extends-fusion-agentic-applications-with-introduction-of-fusion-claw-2026-09-29/

Oracle Corporation (2026k) Oracle Fusion Cloud Service Descriptions, v092926. Contract document, effective September 29, 211 pages, Table 1 “Action Value Factors per Action Type” read by a human on October 6, 2026. Available at: https://www.oracle.com/contracts/docs/oracle-fusion-cloud-service-desc-1843611.pdf

Oracle Corporation (n.d.a) Oracle AI World, October 25 to 28, 2026, Las Vegas. Event page. Available at: https://www.oracle.com/ai-world/

Oracle Corporation (n.d.b) Oracle AI for Fusion Applications. Product page, updated August 4, 2026. Available at: https://www.oracle.com/applications/fusion-ai/

Oracle Corporation (n.d.c) Executive Leadership, Mike Sicilia, Chris Leone. Corporate pages. Available at: https://www.oracle.com/corporate/executives/

Oracle documentation, contract pages and repository

Oracle Help Center (2026a) Manage AI unit budget, consumption, and controls. Release 26D Common Technologies and User Experience What’s New, feature f51348, updated September 21. Available at: https://docs.oracle.com/en/cloud/saas/readiness/common/26d/common26d/26D-common-wn-f51348.htm

Oracle Help Center (2026b) Configure an AI guidance node in a workflow. Release 26D Common Technologies and User Experience What’s New, feature f51676, delivered September 21. Available at: https://docs.oracle.com/en/cloud/saas/readiness/common/26d/common26d/26D-common-wn-f51676.htm

Oracle Help Center (2026c) How do I configure the retention period for generative AI data? Oracle AI for Fusion Applications Questions and Answers, build stamp September 22. Available at: https://docs.oracle.com/en/cloud/saas/fusion-ai/aiaqa/how-do-i-configure-the-retention-period-for-generative-ai-data.html

Oracle Help Center (2026d) Where are the locations of the OpenAI endpoints for AI agents in Fusion Applications? Oracle AI for Fusion Applications Questions and Answers, document G22516-32, build stamp September 22. Available at: https://docs.oracle.com/en/cloud/saas/fusion-ai/aiaqa/where-are-the-locations-of-the-openai-endpoints-for-ai-agents-in-fusion-applications.html

Oracle Help Center (2026e) How do I use AI Agent Studio? Release 26D guide, document G59779-01, September 24. Available at: https://docs.oracle.com/en/cloud/saas/fusion-ai/26d/aiaas/toc.htm

Oracle Help Center (2026f) Add a Policy Model to a Policy Node of Workflows. Release 26D guide, September 24. Available at: https://docs.oracle.com/en/cloud/saas/fusion-ai/26d/aiaas/add-a-policy-model-to-a-policy-node-of-workflows.html

Oracle Help Center (2026g) Configure an AI Guidance Node in a Workflow. Release 26D guide, September 24. Available at: https://docs.oracle.com/en/cloud/saas/fusion-ai/26d/aiaas/configure-an-ai-guidance-node-in-a-workflow.html

Oracle Help Center (2026h) Large Language Models (LLMs) in AI Agent Studio. Release 26D guide, September 24. Available at: https://docs.oracle.com/en/cloud/saas/fusion-ai/26d/aiaas/large-language-models-llms-in-ai-agent-studio.html

Oracle Help Center (2026i) Regional Availability for Generative AI Models. Oracle Cloud Infrastructure documentation, updated September 24. Available at: https://docs.oracle.com/en-us/iaas/Content/generative-ai/model-endpoint-regions.htm

Oracle Help Center (2026j) Use Oracle Fusion Claw to create autonomous outcomes with agentic applications. Release 26D Common Technologies and User Experience What’s New, feature f51374, document G61091-02, delivered September 28. Available at: https://docs.oracle.com/en/cloud/saas/readiness/common/26d/common26d/26D-common-wn-f51374.htm

Oracle Help Center (2026k) Oracle AI for Fusion Applications: Fusion Claw Capabilities. Documentation page, created October 2. Available at: https://docs.oracle.com/en/cloud/saas/fusion-ai/fusion-claw.html

Oracle Help Center (n.d.) Oracle Cloud Applications Readiness, ERP, HCM and SCM. Readiness pages, fetched October 5, 2026. Available at: https://docs.oracle.com/en/cloud/saas/readiness/erp.html , https://docs.oracle.com/en/cloud/saas/readiness/hcm.html and https://docs.oracle.com/en/cloud/saas/readiness/scm.html

Oracle (2026) README.md, oracle/fusion-ai-studio, main and release-26C branches. GitHub repository, changelog to August 21. Available at: https://github.com/oracle/fusion-ai-studio

Oracle blogs

Oracle Fusion Development (2025) Quarterly updates made easy. The Fusion Insider blog, December 10. Available at: https://blogs.oracle.com/fusioninsider/quarterly-updates-made-easy

Oracle Fusion Development (2026a) A playbook for scaling business value with Fusion AI. The Fusion Insider blog, August 27. Available at: https://blogs.oracle.com/fusioninsider/a-playbook-for-scaling-business-value-with-fusion-ai

Oracle Fusion Development (2026b) Announcing AI Studio Skill—pro-code tools for developers. The Fusion Insider blog, August 27. Available at: https://blogs.oracle.com/fusioninsider/announcing-ai-studio-skill-pro-code-tools-for-developers

Oracle Fusion Development (2026c) Fusion Apps experts recommend these AI World sessions. The Fusion Insider blog, September 16. Available at: https://blogs.oracle.com/fusioninsider/fusion-apps-experts-recommend-these-ai-world-sessions

Rajgarhia, G. (2026) Policy Model in Workflow Agents. Oracle Fusion Development Center of Excellence blog, August 5. Available at: https://blogs.oracle.com/fusioncoe/policy-model-in-workflow-agents

Rajgarhia, G., Satyamurthy, A., Ragul, A., Mathew, R. and Kerr, B. (2026) Identity and Security in AI Agent Studio. Oracle Fusion Development Center of Excellence blog, July 27. Available at: https://blogs.oracle.com/fusioncoe/identity-and-security-in-ai-agent-studio

Disclosed briefing and review material, not public

Oracle Corporation (2026l) Fusion Claw press and analyst deck, 23 slides, shown to analysts under embargo on September 25. Not public.

Oracle Corporation (2026m) Analyst pre-briefing on Fusion Claw, given by Kaushal Kurapati, Group Vice President, Fusion AI, Oracle Applications and AI, on September 25. Not public. Identified in this research report as a disclosed evidence source.

Oracle Corporation (2026n) Written response to Greyhound Research’s factual review of this research report, October 12. Not public.

Named media carrying Oracle executives

Dey, V. (2026) Oracle Fusion Claw Adapts OpenClaw’s Agentic AI Approach For Enterprises. Forbes, September 29. Available at: https://www.forbes.com/sites/victordey/2026/09/29/oracle-fusion-claw-adapts-openclaws-agentic-ai-approach-for-enterprises/

Essex, D. (2026) Oracle Fusion Claw the new grab for fully autonomous agentic AI in ERP. TechTarget, September 29. Available at: https://www.techtarget.com/it-strategy/news/366651382/Oracle-Fusion-Claw-the-new-grab-for-fully-autonomous-agentic-AI-in-ERP

Ghoshal, A. (2026) Oracle Fusion Claw pinches AI costs, tightens grip on policies. CIO, September 29. Available at: https://www.cio.com/article/4228371/oracle-fusion-claw-pinches-ai-costs-tightens-grip-on-policies.html

Gillin, P. (2026) Oracle expands AI agent features with Fusion Claw. SiliconANGLE, September 29. Available at: https://siliconangle.com/2026/09/29/oracle-expands-ai-agent-functionality-with-fusion-claw/

Lauchlan, S. (2026) Unhooding Claw – Oracle’s latest Agentic Applications extension snaps into place. Diginomica, September 29. Available at: https://diginomica.com/unhooding-claw-oracles-latest-agentic-applications-extension-snaps-place

Singh, P. (2026) Oracle’s AI agents can now execute enterprise work — here’s what CIOs should know. CIO&Leader, July 15. Available at: https://www.cioandleader.com/oracles-ai-agents-can-now-execute-enterprise-work-accountability-is-the-question-cios-should-ask-next/

Yeung, K. (2026) With Fusion Claw, Oracle Is Betting Businesses Will Let AI Act Without Asking First. The AI Economy, September 29. Available at: https://theaieconomy.substack.com/p/oracle-fusion-claw-ai-acts-without-asking

Category vendor documentation reviewed

Amazon Web Services (2026) Amazon Bedrock AgentCore harness is now generally available, with the AgentCore developer guide pages Security and access controls and Release notes. Blog and documentation, June 18 and September. Available at: https://aws.amazon.com/blogs/machine-learning/amazon-bedrock-agentcore-harness-is-now-generally-available-go-from-idea-to-production-grade-agent-in-minutes/ , https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/harness-security.html and https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/release-notes.html

Microsoft (2026a) Build 2026 news. GitHub microsoft/Build26-news, June 2. Available at: https://github.com/microsoft/Build26-news/blob/main/news.md

Microsoft (2026b) Governing Agent Identities, Microsoft Entra ID Governance. Microsoft Learn, updated June 16. Available at: https://learn.microsoft.com/en-us/entra/id-governance/agent-id-governance-overview

Microsoft (2026c) View Copilot Studio audit logs in Microsoft Purview. Microsoft Learn, August 18. Available at: https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-logging-copilot-studio

OpenAI (2026) Introducing the Agents API. Announcement, September 10. Available at: https://openai.com/index/introducing-the-agents-api/

ServiceNow (2026) What’s new in AI Control Tower for August & September 2026. ServiceNow Community, September 14. Available at: https://www.servicenow.com/community/ai-control-tower-articles/what-s-new-in-ai-control-tower-for-august-amp-september-2026/ta-p/3597749

UiPath (2026) Maestro, September 2026 release notes. Documentation, September 17. Available at: https://docs.uipath.com/maestro/automation-cloud/latest/release-notes/september-2026

OpenClaw, the naming convention, advisories and trademark records

Automation Anywhere (2026) Automation Anywhere Collaborates with Cisco, NVIDIA, Okta, and OpenAI, Launching EnterpriseClaw to Run Next-Generation AI Agents Inside Enterprise Systems. Press release, May 19. Available at: https://www.automationanywhere.com/company/press-room/automation-anywhere-collaborates-cisco-nvidia-okta-and-openai-launching

Cisco (2026) Cisco Reimagines Security for the Agentic Workforce. Press release, March 23. Available at: https://newsroom.cisco.com/c/r/newsroom/en/us/a/y2026/m03/cisco-reimagines-security-for-the-agentic-workforce.html

Cyber Security Agency of Singapore (2026) Advisory on Cybersecurity Risks of OpenClaw, AD-2026-005. May 28. Available at: https://www.csa.gov.sg/alerts-and-advisories/advisories/ad-2026-005/

NVIDIA (n.d.) Build a Claw. Page, fetched October 6, 2026. Available at: https://www.nvidia.com/en-us/ai/build-a-claw/

OpenClaw Foundation (2026) Blog index, January 29 to October 2, 2026, including OpenClaw Enterprise, September 29. Available at: https://openclaw.ai/blog

United States Patent and Trademark Office (2026) Trademark Status and Document Retrieval, serial 99372328, FUSION CLAW, Wilcox Industries Corp., Class 009, published February 17, Notice of Allowance April 14. Available at: https://tsdr.uspto.gov/statusview/sn99372328

Wilcox Industries Corp. (2024) Wilcox Details FUSION CLAW and RAPTAR Xe Events for SHOT SHOW 2024. Company news, January 16. Available at: https://wilcoxind.com/news/articles/wilcox-details-fusion-claw-and-raptar-xe-events-for-shot-show-2024

Regulator, government and standard-setter texts

Board of Governors of the Federal Reserve System, Office of the Comptroller of the Currency and Federal Deposit Insurance Corporation (2026) SR 26-2, Revised Guidance on Model Risk Management, attachment. April 17. Available at: https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm

California Privacy Protection Agency (2025) Regulations on cybersecurity audits, risk assessments and automated decisionmaking technology, approved text, and announcement of September 23, 2025, with automated decisionmaking requirements applying from January 1, 2027. Available at: https://cppa.ca.gov/regulations/pdf/ccpa_updates_cyber_risk_admt_appr_text.pdf and https://cppa.ca.gov/announcements/2025/20250923.html

Colorado Attorney General (2026) ADMT & Chatbot Pre-Rulemaking Considerations Document, June, and the artificial intelligence rulemaking page, law effective January 1, 2027. Available at: https://coag.gov/app/uploads/2026/06/ADMT-Chatbot-Pre-Rulemaking-Considerations-Document.pdf and https://coag.gov/ai/

European Commission (2026) Commission opens second-phase consultation on Quality Jobs Act. Directorate-General for Employment, Social Affairs and Inclusion, July 20. Available at: https://employment-social-affairs.ec.europa.eu/news/commission-opens-second-phase-consultation-quality-jobs-act-2026-07-20_en

European Parliament (2025) Digitalisation, artificial intelligence and algorithmic management in the workplace – shaping the future of work. Resolution of December 17, CELEX 52025IP0337. Available at: https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX%3A52025IP0337

European Union (2024) Regulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 12, 14 and 26 and Annex III, as amended by Regulation (EU) 2026/1744, applying to Annex III systems from December 2, 2027. Text as published by the AI Act Explorer. Available at: https://artificialintelligenceact.eu/article/14/

European Union (2026) Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 (Digital Omnibus on AI). EUR-Lex, in force July 27. Available at: https://eur-lex.europa.eu/eli/reg/2026/1744/oj

Financial Conduct Authority (2026) The Mills Review: AI and the future of retail financial services. July. Available at: https://www.fca.org.uk/publication/corporate/the-mills-review.pdf

Infocomm Media Development Authority (2026) Model AI Governance Framework for Agentic AI, Version 1.0. January 22. Available at: https://www.imda.gov.sg/-/media/imda/files/about/emerging-tech-and-research/artificial-intelligence/mgf-for-agentic-ai.pdf

International Auditing and Assurance Standards Board (2026) Proposed International Standard on Auditing 500 (Revised), Audit Evidence, exposure draft. August, comments due December 15. Available at: https://ifacweb.blob.core.windows.net/publicfiles/2026-08/IAASB-Audit-Evidence-Risk-Response-ISA-500-Exposure-Draft.pdf

Jain, R. (2026) From Digital Banking to Resilient Banking – Technology, Cyber Security and AI as Pillars of Trust. Speech, Reserve Bank of India, SBI Banking and Economic Conclave, Mumbai, September 24. Available at: https://www.rbi.org.in/Scripts/BS_SpeechesView.aspx?Id=1581

Malhotra, S. (2026) Preserving Financial Stability in an Evolving World. Speech, Reserve Bank of India, Fifth Kautilya Economic Conclave, October 3. Available at: https://www.rbi.org.in/Scripts/BS_SpeechesView.aspx?Id=1582

Murmu, S. C. (2026) Trusted Innovation: Shaping the Future of Finance. Speech, Reserve Bank of India, Global Fintech Fest, Mumbai, September 11. Available at: https://www.rbi.org.in/Scripts/BS_SpeechesView.aspx?Id=1579

Office of the Privacy Commissioner, New Zealand (n.d.) Principle 3A and Principle 8, Privacy Act 2020, with Ministry of Justice, New Zealand (2026) New privacy information principle now in force, IPP 3A in force May 1, 2026. Available at: https://www.privacy.org.nz/privacy-act-2020/privacy-principles/3a/ , https://www.privacy.org.nz/privacy-act-2020/privacy-principles/8/ and https://www.justice.govt.nz/about/news-and-media/news/new-privacy-information-principle-in/

Office of the Australian Information Commissioner (2026) New resources on transparency for use of AI and automated decision-making, and APP 1.7-1.9 Transparency Obligation fact sheet. September 30. Available at: https://www.oaic.gov.au/news/media-centre/new-resources-on-transparency-for-use-of-ai-and-automated-decision-making

Outlook Money (2026) Sebi to Issue AI/ML Guidelines for Capital Markets: Mandatory ‘Kill Switch’ and Human Oversight Framework Explained. Named media report of a regulator’s statement of August 19, updated August 20. Available at: https://www.outlookmoney.com/invest/sebi-to-issue-aiml-guidelines-for-capital-markets-mandatory-kill-switch-and-human-oversight-framework-explained

Press Information Bureau, Government of India (2025) Government notifies DPDP Rules to empower citizens and protect privacy. November 14. Available at: https://www.pib.gov.in/PressReleasePage.aspx?PRID=2190014&reg=3&lang=2

Reserve Bank of India (2026) RBI issues draft Guidance on Regulatory Principles for Model Risk Management. Press release 2026-2027/528, June 24. Available at: https://www.rbi.org.in/scripts/BS_PressReleaseDisplay.aspx?prid=63006

United Kingdom (2025) Data (Use and Access) Act 2025, section 80, automated decision-making provisions in force February 5, 2026. legislation.gov.uk. Available at: https://www.legislation.gov.uk/ukpga/2025/18/section/80

United States General Services Administration, Office of the Chief Information Security Officer (2023) IT Security Procedural Guide: Robotic Process Automation (RPA) Security, CIO-IT Security-19-97, Revision 3. February 14. Available at: https://www.gsa.gov/system/files/Robotic-Process-Automation-(RPA)-Security-[CIO-IT-Security-19-97-Rev-3]-02-14-2023.pdf

Employee representatives

FNV, ARBO netwerk (2025) Medezeggenschap en AI. Guidance for works councils, December 13. Available at: https://www.fnv.nl/getmedia/8b24119f-4f18-43e9-9d4f-6a3d08a21c59/AI-in-medezeggenschap-20251203.pdf

Named buyers, industry bodies and surveys

Asia Securities Industry and Financial Markets Association (2026) ASIFMA Response to MAS Consultation Paper on Proposed Guidelines on Artificial Intelligence Risk Management for Financial Institutions. January 30. Available at: https://www.asifma.org/wp-content/uploads/2026/02/2026-01-31-asifma-response-to-mas-ai-risk-mgmt-guidelines-cp.pdf

Crozier, R. (2026) ATO culture keeps a cap on agentic AI ambitions. iTnews, September 23. Available at: https://www.itnews.com.au/news/ato-culture-keeps-a-cap-on-agentic-ai-ambitions-629118

Estrada, S. (2026) Dell’s CFO is using AI agents to run his finance team—and has helped the AI business go from $0 to $25 billion. Fortune, March 30. Available at: https://fortune.com/2026/03/30/dells-cfo-is-using-ai-agents-to-run-his-finance-team-and-has-helped-the-ai-business-go-from-0-to-25-billion/

ISACA (2026) New ISACA Research: 56% of Digital Trust Pros Don’t Know How Fast They Could Shut Down AI After a Security Incident. Press release via Business Wire, March 23, a global pulse poll of more than 3,400 digital trust professionals. Available at: https://www.businesswire.com/news/home/20260323090694/en/New-ISACA-Research-56-of-Digital-Trust-Pros-Dont-Know-How-Fast-They-Could-Shut-Down-AI-After-a-Security-Incident

UK Finance (2025) Agentic AI and Regulatory Challenges, response to the Digital Regulation Cooperation Forum. November. Available at: https://www.ukfinance.org.uk/system/files/2025-11/UK%20Finance%20response%20to%20DRCF%20consultation.pdf

Implementation partners and practitioners

Mead, K. (2026) Oracle AI Agents: Understanding the pricing model before you commit. Version 1 blog, June 4. Available at: https://www.version1.com/blog/oracle-ai-agents-understanding-the-pricing-model-before-you-commit/

Paliwal, A. (2026) Oracle Fusion AI Pricing Explained: AI Units, Agentic Apps Costs in 26C. Practitioner newsletter, June 4. Available at: https://arpitpaliwaloracleaierp.substack.com/p/oracle-fusion-ai-pricing-explained

SMACT Works (2026) Oracle Fusion Claw: Is Your Organization Ready for Agentic AI? Partner blog, September 30. Available at: https://www.smactworks.com/insights/blogs/2026/09/oracle-fusion-claw-agentic-ai.html

Uppada, S. (2026) Oracle Fusion AI Agent Licensing: Production Decisions That Matter. Calfus blog, June 19. Available at: https://www.calfus.com/resources/blog/oracle-fusion-ai-agent-licensing-production-decisions-that-matter/

Greyhound Research prior writing

Gogia, S. V. (2026) Everyone Is Building AI Agents. Oracle Is Moving The Contest Inside The Runtime. Greyhound Research research report, July 20. Available at: https://greyhoundresearch.com/everyone-is-building-ai-agents-oracle-is-moving-the-contest-inside-the-runtime/

↑ Back to the section guide

Disclosures: Independence, Authoring, And Copyright

This research report is current through October 6, 2026, and reflects Oracle’s written response of October 12, 2026 to its factual review where the text says so.

Independence Disclosure

Research under the Greyhound Research brand is independent, and no organization holds a right of approval over its content or publication. Greyhound Consilium is the B2B thought-leadership arm of The House Of Greyhound and a sister brand to Greyhound Research. It builds on Greyhound Research’s independent research through collaborations that produce thought-leadership content, research, and practical toolkits for decision-makers. Every Greyhound Consilium publication names its partner on the face of the document. This research report is published under the Greyhound Research brand.

Greyhound Research’s analysis rests on two bodies of evidence: its own primary research, and source evidence verified against the document or briefing that supports each claim. The primary research comprises sustained coverage of the enterprise technology market, Greyhound Fieldnotes from advisory engagements with buy-side clients, and the Greyhound CIO Pulse editions cited by name and date in the text.

For this research report, the source evidence comprises Oracle’s public documents, a pre-briefing and deck provided by Oracle on September 25, 2026, Oracle’s written response to its factual review, received on October 12, 2026, and the public record as re-verified on October 5 and 6, 2026. Greyhound Research provided this research report to Oracle before publication and invited comment on factual accuracy and supporting evidence only. Greyhound Research verified each proposed correction independently, assessed its implications for the analysis, conclusions and recommendations, and made changes only where the evidence warranted them. Where the text relies on Oracle’s written response, it attributes the statement to Oracle. Editorial control remained entirely with Greyhound Research throughout.

Greyhound Research undertakes advisory and consulting engagements with enterprise buyers and with technology providers. No such engagement confers influence over research findings, and no commercial relationship determines the content of a research report or the decision to publish it.

Greyhound Research discloses any financial interest in, or current commercial relationship with, an organization named in a research report. Oracle is a client of Greyhound Research. The engagement covers only thought-leadership content published under the Greyhound Consilium brand and did not fund this research report. Greyhound Research holds no other financial interest in Oracle.

Authoring Disclosure

This material reflects the analysis, viewpoint and editorial judgment of Greyhound Research. The position presented is developed from Greyhound Research’s research, advisory work and accumulated institutional knowledge. Depending on the assignment, AI may be used to retrieve relevant material; structure, draft or polish prose from analyst-provided concepts, dictation, notes, rough drafts and source material; support secondary research and factual cross-checking; and assist with editing, condensation and consistency with Greyhound Research’s house style. AI does not independently establish Greyhound Research’s viewpoint, analytical position or conclusions. Material factual claims are verified against source evidence before release. Greyhound Research reviews and approves the final material and retains responsibility for it.

Copyright Policy

All content in this document and on the Greyhound Research website is protected by copyright law and may not be reproduced, distributed, transmitted, displayed, published, or broadcast without the prior written permission of Greyhound Research or, in the case of third-party materials, the prior written consent of the copyright owner of that content. You may not alter, delete, obscure, or conceal any trademark, copyright, or other notice appearing in any Greyhound Research content. Quotation of up to one hundred words is permitted where it carries attribution to Greyhound Research and a link to the published page. Reproduction, redistribution or republication beyond that extent, in whole or in part, by email or in any medium including websites, newsletters and intranets, requires prior written permission, which may be requested at the address below. The published version of each content piece on the Greyhound Research website carries sharing tools for readers who wish to share it.

For questions arising from these disclosures, including licensing requests and media inquiries, contact the Greyhound Research Community Relations Team at connect@thofgr.com.

↑ Back to the section guide